ConvFormer: Revisiting Transformer for Sequential User Modeling
Published in arXiv preprint, 2023
Recommended citation: Hao Wang, Jianxun Lian, Mingqi Wu, Haoxuan Li, Jiajun Fan, Wanyue Xu, Chaozhuo Li, Xing Xie. "ConvFormer: Revisiting Transformer for Sequential User Modeling." arXiv preprint arXiv:2308.02925, 2023. https://arxiv.org/abs/2308.02925
ConvFormer revisits Transformer-style models for sequential user modeling, the recommender task of predicting the next item a user will want. An analysis of the item-to-item mechanism yields three criteria for efficient sequential user models; ConvFormer, a simple modification of the Transformer that meets them, reaches state-of-the-art results, and an acceleration technique keeps it efficient on long behavior sequences.